Any-Order Generation for Pretrained Autoregressive Image Models
October 28, 2026
A new method enables inpainting and editing in pretrained raster-scan autoregressive models without the 3x compute cost of training on multiple token orderings. This allows existing AR models to express their learned joint distribution through non-sequential generation.
HOW THIS AFFECTS YOU
●
researcherYou can bypass expensive multi-order training for autoregressive architectures.
●
designerThis unlocks more flexible image editing capabilities in existing AR-based tools.